Fig. 7
From: Marigold: a machine learning-based web app for zebrafish pose tracking
Marigold’s WebAssembly-based neural network implementation achieves CPU performance competitive with that of PyTorch. A Training speeds for the touch-evoked response dataset. B Training speeds for the visuomotor response dataset. C Inference speeds for the touch-evoked response dataset. D Inference speeds for the visuomotor response dataset. Measurements reflect single-threaded CPU performance. Training speeds were measured using gradient accumulation with a batch size of 1 to obtain an effective batch size of 16. Inference speeds were measured with a batch size of 1. Data represent the mean of 10 independent measurements